Your Choice Comes Pre-Selected

France's competition authority, the Autorité de la concurrence, has put three companies under scrutiny for market dominance. In an opinion published on 17 July, running to 3,700 pages with its annexes, it estimates that OpenAI, Google, and Anthropic together control more than 84% of the AI agent market, and warns that the move from chatbots to autonomous agents could tighten the digital economy around a handful of gateways.

Eighty-four percent in three hands is the headline. What that share controls is the story. The models get copied, and the price per token is collapsing. The prize is the default, the choice that fires when you stop choosing. Ask an agent to "find me a toaster under fifty euros" and three seconds later your cart is full. You never saw the two models it left out. Your choice arrived pre-selected, on criteria nobody showed you.

It is worth seeing how the regulator got there, because it is the part practitioners should steal. It did not ask the companies how neutral their agents were. It built its own, put 550 shopping questions to them, and logged which sites they actually visited and cited. It instrumented the system and measured its behavior instead of trusting the description. That is the difference between reading a service's documentation and watching it run in production: the first tells you what should happen, the second tells you what does.

What happens is that the agent becomes a default, and a default is quiet power. It is the pre-checked box, the first result, the answer that lands before you frame the alternative. Whoever owns the default owns the outcome without having to convince anyone. It is access control applied to desire: you keep the right to choose and lose the occasion to exercise it, because the decision already happened upstream, inside a system answering to its own incentives. I have called agents a permission that changes hands. The permission at stake here is the one to decide for you.

The economic word for this is disintermediation. Retail's fight has left the old axis of Walmart against Target for a new one: who becomes the single point of entry, ChatGPT, Gemini, Perplexity, Amazon Rufus. Sellers who launched their own agent grew faster, and in doing so handed the agent the thing that matters, the customer relationship and the data that describes it. In 2025, AI influenced 20% of global online sales, 262 billion dollars. That flow no longer passes through the seller's storefront. It passes through the default of three companies.

An agent that chooses is convenient, and convenience is no sin. Nobody misses comparing twelve toasters by hand. Delegating the drudgery of a trivial purchase is exactly what the machine is for. Delegating is fine. The trouble is where it stops.

Now the reverse, which is the part that counts. Pre-selection is a design decision, and a design decision can be made the other way. The same technology splits into two tools that look identical. An agent can sedate you or sharpen you. It sedates you when it returns the result and hides the reasoning: you take the toaster and never learn why that one. It sharpens you when it shows the criteria, the models it dropped, the conflict of interest behind the ranking, and leaves you the last word. The first saves you three seconds and shaves a gram of judgment each time, until you notice you have stopped choosing. The second extends how far your attention reaches: it reads twelve spec sheets for you and brings back the one you could not have found alone. In that second form the machine does the work of a key. It opens a door you already meant to open, faster.

Here is why 84% matters. Whether an agent sedates or sharpens you depends on who owns the default. A company that lives on commissions has no incentive to show you the option it earns nothing from. When the default that chooses for a continent sits in three hands, the agent's design stops being your choice as a consumer and becomes their choice as a product. The version that sharpens you is within technical reach. The incentive to build it is not.

For anyone who works, the stake is exactly this. Every professional is learning to delegate pieces of judgment to an agent: which supplier, which draft, which candidate to move to the interview. Ask one thing of it: does it hand you the why, or only the what? An assistant that brings you the conclusion makes you faster today and poorer tomorrow, because the muscle you stop training wastes away. One that brings you the conclusion with the work underneath leaves the verb in your hands. Your choice comes pre-selected only as long as you use the tool that chooses for you. There is one that makes you choose. Use that one, while the choice is still yours to make.